Facial Expression And Gender Recognition Based On Convolutional Neural Networks

Human face involves lots of information, including gender, facial expression, identity, age and so on. Under the urgent drive of the development of the field of electronic information security, such as public social security, economic property security, military criminal investigation on terrorism, human computer interaction and so on, facial expression and gender recognition technology have already been a kind of advanced technology with great potential for development and current research hotspot in the field of computer vision. It is of significant application value to make the computer have the human intelligence and substitute human to remember, identify and realize the real intelligent age. However, facial expression and gender recognition technology are also the difficult point in the field of computer vision. It is mainly because the environmental factors like lights, expression, post and screen as well as the shooting behaviors have influence on the process of acquiring face images. Thus, an excellent facial expression and gender recognition calculation should not be sensitive to those factors.Convolutional Neural Networks are a kind of new neural networks, which connect traditional artificial neural networks with deep learning technology. They are characteristic of local receptive field, hierarchical structure and global training connecting feature extraction with classification process, which obtained great success in the field of image process. Convolutional Neural Networks mainly have two features, which the first one is adopting local join strategy in neuron and the second one is the weight sharing of neuron on the same level. The network structure adopted local join and weight sharing reduces the complexity of the model itself and number of parameters needed to be trained. This kind of network structure can obtain translation, scale and deformation invariance to a certain extent.The main studies of this thesis are as follows:1.At first, this thesis systematically introduces the research status of facial expression recognition and gender recognition both at home and abroad, summarizes the origin of deep learning and its research findings are also discussed in this thesis. After that, this thesis introduces the development process of Neural Network, in which the algorithm principle and classic network structure of Convolution Neural Networks are the focus.2.This thesis focuses on facial expression and gender recognition. Firstly, as for facial expression, this thesis amends the network structure of AlexNet, designs a new Convolution Neural Networks in accordance with the features of facial expression task, then adds a Batch Normalization layer. By doing this, the accuracy rate can increase by 3%. Secondly, according to the samples of data set adopted by facial expression recognition, this thesis uses the methods of fine tune. Comparing with original training,the accuracy can indeed be improved by 2% in GoogLeNet. Moreover, when using fine tuning with VGGNet, the accuracy rate can also increase to 71.27% at most, which proves that the methods of fine tuning is better than retraining in terms of data sets in this thesis. When by using the GoogLeNet to conduct the fine tuning experiment, this thesis also makes a comparison of the performance between the Hinge loss function and Softmax loss function and finds that the latter is better than the former. Finally, aiming at the latest research tendency, this thesis designs multiple networks to fuse. Finally, it turns out that the performance of multinetwork is worse than that of a single network on the basis of smaller data sets. Moreover, as for gender recognition, a new Convolution Neural Networks is invented on the basis of VGGNet. It turs out that the accuracy has increased to 90.82% on the adience data set and on the mygender data set, which accuracy is 97.10%. Finally, it is amazing that the accuracy is up to 99.44% when by using the VGGNet to conduct the fine tuning experiment on the basis of mygender data set.3.By using the Convolution Neural Networks model of facial expression together with gender data set and other instruments such as dilb and Caffe, we could establish facial expression and gender recognition system on Windows 7. The system can be used to detect recognition real-timely and accurately as well as make a preemptive arrangement for the following work.